Our customer wanted to know beforehand which clients were about to churn. Using ML & AI, we managed to succesfully predict these customers.
AI-Based Predictive Analytics
A large issue faced by many financial institutions is customer attrition, the loss of customers by a business. Customer attrition has been a massive problem not only for the financial industry but more many other businesses. A lot of future plans and projects are designed with the assumption that their customers will stay for the long run.
This is why our clients wanted a way to predict which customers were on the brink of churning, so that they can target those customers before they leave. By forecasting which customers are about to leave, they can pay more attention to those customers with the hope that they will stay. To make such predictions, a complex Machine Learning model which could perform predictive analytics was required.
To tackle this challenge, we first collected as much historical data as possible from the client as well as other reliable data sources such as google analytics. We realised that the more data we had about the customers, the easier it would be to figure out which factors were important and should be fed into our Machine Learnng model. Some of the data we included was customer demographics, customer profiles, transactions, payment behaviour, call centre data, and engagement with mobile apps. By the end of this process, we had over hundreds thousands of data points in over 200 variables.
With 200 variables, it was essential to narrow down to the most important parameters and then build different models with different combination of variables. Feeding a Machine Learning model with too many parameters can make it either less accurate, or more slow in its predictions. Alongside financial expertise of our clients and our dedicated Artificial Intelligence team, we chose which parameters would make the most sense.
However, even with narrowing down the amount of variables required to feed the system, we tried different combinations. We needed to build a model which worked upto the client's satisfaction. We ended up building 28 unique models trying different Machine Learning techniques to finally get a system which got the accuracy the client wanted.
In order to test our model, it was put to test against live data that was coming each minute and it was able to successfully predict which customers were about to churn to a level our clients were happy.